Agent00HL7

The 21st Century Cures Act gave patients much greater access to their health records — but radiology reports are full of difficult language and medical jargon, which can lead to misinterpretation and anxiety.

• ▪AI-generated letters that explain a radiology report in plain language while preserving its medical context and accuracy
• ▪An accuracy check that matches ICD-10 codes between the original report and the letter
• ▪A readability check using the Flesch-Kincaid metric
• ▪A workflow that reduces how much proofreading medical professionals need to do

• ▪Instead of one zero-shot prompt, the letter goes through an iterative self-refinement loop based on the Reflexion framework for AI agents
• ▪Each draft is checked for accuracy (do the letter's ICD-10 codes match the report's?) and readability (Flesch-Kincaid)
• ▪The loop is programmatic, so it improves accuracy while minimizing the need for human input

• ▪Tested on 20 randomized radiology reports
• ▪94.94% ICD-10 verification accuracy with the multi-agent approach, vs. 68.23% with zero-shot prompting
• ▪81.25% of final letters needed no corrections for accuracy or readability, vs. 25% of zero-shot letters
• ▪2nd place at the SIIM Hackathon